On Mon, Feb 05, 2007, Alex Rousskov wrote: > Hi there, > > While applying the BodyPipe changes to ftp.cc and http.cc, I would like > to move lots of common code into ServerStateData, the class that both > FTP and HTTP "servers" already share. > > One of the differences I noticed is how FTP and HTTP write request body > pieces when receiving them from the client_side. Before doing each > write, FTP calls commSetSelect(COMM_SELECT_WRITE) in > ftpDataWriteCallback(). HTTP appears to simply call comm_write, without > bothering to wait for the file descriptor to be selected for writing. > > To me, it looks like http.cc code (and client_side code as well) are > using "optimistic" writes. They rely on comm_* code to handle situations > where the descriptor is not ready. Ftp.cc seems to be more cautious, but > I do not know why.
Its a hold-over to how things were done in the past. Please try to take advantage of comm_write() where appropriate as it'll make things easier to port in the long run. (Heck, Squid-2 suffers from the same split-brain personality when it comes to network IO. I'm pretty sure comm_write() was there far before I got to the networking layer.) > Is the difference between ftp.cc and http.cc essential? Can I remove > commSetSelect(COMM_SELECT_WRITE) logic from the FTP code and use bare > comm_write calls instead? Yup. > Alex. > P.S. FWIW, the BodyPipe work is progressing better than I expected. I > was even able to remove a large chunk of ICAP code, and the result may > be easier to comprehend and tweak. The server side (http.cc and ftp.cc) > code is the most complex one as it has to deal with up to three(!) Nifty! Its good to see squid-3 work progressing. Adrian
